AI Potluck
Back to Gap Map Model components / Multimodal models

Holo

H Company
open weights / Overall score: 3.7

Holo is H Company's family of vision-language models for computer-use agents, fine-tuned from Qwen bases to read screens and act on web, desktop and mobile interfaces. The current Holo3.1 release comes in 0.8B, 4B, 9B and 35B-A3B sizes with native function calling and FP8, NVFP4 and GGUF builds. H Company's larger Holo3-122B model is available only through its API.

Scored on Holo3.1. Some earlier Holo sizes carry a non-commercial or Qwen Research license rather than Apache 2.0.

Openness

3 high confidence
3.0
weights
open(safetensors on the Hub, ungated)
data
described(the Holo3 card names a mix of open datasets, synthetic trajectories and human-annotated samples
code
closed(no code repository
license
Apache-2.0(OSI

Every Holo3.1 checkpoint is released under Apache 2.0. H Company describes its training mix of synthetic navigation data and reinforcement learning but publishes neither the data nor any training code.

Adoption

3 medium confidence
3.0

Hugging Face downloads over the trailing 30 days, summed across H Company's Holo1 to Holo3.1 checkpoints, including its own quantized builds. Most of it is the GGUF build of the 35B model, and use of the hosted API is not counted.

Capability

4 medium confidence
4.0

Holo is built to operate computers: it reads screenshots and returns actions or function calls on web, desktop and mobile, with strong vendor-reported AndroidWorld and OSWorld-Verified scores. It is level with UI-TARS, and like it takes no audio input.

Verified 2026-09-26